💻 computer science

Knowledge-Constrained Shape Optimization with a Mixture-of-Experts Neural Operator for High-Confidence Design

This paper proposes a knowledge-constrained shape optimization framework that integrates a Mixture-of-Experts Neural Operator with uncertainty-aware feedback to automate engineering constraints and achieve high-confidence drag reductions of 4–10% across diverse vehicle geometries.

Original authors: Wenhao Fan, Yuanwei Bin, Jianghan Gu, Wenfa Luo, Jiao Xiang, Yuntian Chen, Shiyi Chen

Published 2026-07-14

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Wenhao Fan, Yuanwei Bin, Jianghan Gu, Wenfa Luo, Jiao Xiang, Yuntian Chen, Shiyi Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a master sculptor trying to carve a faster, sleeker car out of a giant block of clay. But there's a catch: you can't just smash the clay anywhere. You have strict rules. You can't change the shape of the windows because people need to see out, you can't make the car too wide or it won't fit in a garage, and you can't mess with the engine bay. In the real world, experienced engineers usually have to sit down and manually draw these "safe zones" on the blueprint before they start tweaking the design. It's slow, and it relies entirely on their memory and experience.

This paper introduces a new, super-smart assistant that does two big things: it automatically draws those "safe zones" for you, and it builds a crystal ball that predicts how fast your car will go without you having to run a million expensive wind tunnel tests.

The "Smart Rules" Assistant
First, the team taught their computer to read a library of engineering rulebooks, safety guidelines, and design standards. When you show it a picture of a car, the computer uses a "vision-language model" (think of it as a robot that can see and read at the same time) to figure out exactly which parts are allowed to change. It might say, "Okay, we can tweak the front bumper and the hood, but the A-pillar (that stick holding up the windshield) must stay put."

Instead of just guessing, the computer translates these rules into invisible 3D boxes around the car. It then uses a mathematical trick called "Direct Free-Form Deformation" to wiggle the clay inside those boxes. This ensures that every new shape the computer invents is physically possible and respects all the safety and manufacturing rules. It's like having a robot that only lets you sculpt the parts of the car that are allowed to move, keeping the rest perfectly rigid.

The "Expert Panel" Crystal Ball
Now, usually, to see if a new car shape is fast, you have to run a massive computer simulation called CFD (Computational Fluid Dynamics). This is like running the car through a virtual wind tunnel. It's incredibly accurate, but it takes so much computing power that you can't test every single idea.

So, engineers usually use a "surrogate model"—a fast, cheap AI that guesses the wind tunnel results. But here's the problem: if the AI hasn't seen a shape like yours before, it might guess wildly wrong. It's like asking a weather forecaster who only knows about summer to predict a blizzard; they might say "sunny" when it's actually snowing.

The authors built a new kind of AI called a Mixture-of-Experts Neural Operator (MoE-NO). Imagine this isn't just one AI, but a panel of different experts. One expert knows about SUVs, another knows about sedans, and another knows about minivans. When you show the AI a new car shape, a "gating network" (the team leader) asks, "Which expert knows this shape best?" and then combines their answers.

The paper shows that this "expert panel" is much better than previous single-AI models. On a test set of 450 car shapes, the new AI predicted the drag (air resistance) with an error of only 1.16%, beating the best previous model which had an error of 1.52%. Even more importantly, it got the trend right 94.34% of the time. This means if one car is faster than another, the AI correctly guesses which one is faster, which is crucial for finding the best design.

The "Trust but Verify" Safety Net
The coolest part is how the system handles uncertainty. The AI knows when it's guessing. It uses a mathematical "uncertainty score" to check if the car shape it's looking at is something it has seen before.

If the AI is confident: It says, "I'm 99% sure this shape is good," and you move on without running the expensive wind tunnel test.
If the AI is nervous: It says, "I've never seen a shape like this, I'm not sure," and it triggers a "physics-solver-in-the-loop" mechanism. This means it quietly runs a few real, high-fidelity wind tunnel tests just for that specific tricky shape to learn from it, then updates its expert panel.

This "uncertainty-gated" approach saved the day in a test case with an SUV. The AI initially thought it found a design that reduced drag by 4.36%, but because the uncertainty score was high, they ran the real test. The real test showed the AI was wrong; the drag reduction was actually only 4.36% (wait, the paper says the initial prediction was off, and after refinement, the verified reduction was 7.96%). After the AI learned from the real test data, it corrected its guess, and the final verified drag reduction jumped to 7.96%. Without this safety net, they might have accepted a bad design or missed a great one.

The "Out-of-Distribution" Challenge
To really test the system, the researchers threw a curveball: a car shape (called "Sedantest") that was completely different from anything in their training library. The initial AI failed miserably, with a prediction error of 14.88% and a negative accuracy score (R² = -18.42), meaning it was guessing randomly.

But the system's "distribution-level gate" caught this. It realized, "Hey, this is a new type of car we haven't seen!" It then generated 150 new local designs around this weird car, ran the real wind tunnel tests on them, and taught a brand-new "expert" to the panel. After this quick training, the AI's accuracy skyrocketed to an error of just 1.84%. It then successfully optimized this new car, reducing its drag by 7.43%.

What the Paper Says It's NOT
The authors are very clear about what this isn't. They aren't saying this solves every design problem. They aren't claiming the AI works perfectly without any human input or historical data. In fact, they argue against the idea of using a single, global AI model for all car types, showing that it leads to errors when the data is mixed. They also note that this method still requires some real wind tunnel tests (CFD) to verify the results, especially when the AI is unsure; it doesn't replace physics entirely, it just makes the process smarter and safer.

The Bottom Line
In these simulations and tests, the team showed that by combining "smart rules" for shaping the car with a "panel of experts" AI that knows when to ask for help, you can design faster cars more reliably. They achieved drag reductions of 10.52% for a minivan and 4.17% for a sedan without needing extra checks, and up to 7.96% for an SUV after a quick safety check. It's a way to make high-tech design less of a gamble and more of a confident, guided journey.

Technical Summary: Knowledge-Constrained Shape Optimization with a Mixture-of-Experts Neural Operator

Problem Statement

Engineering shape optimization faces dual challenges at the application and technical levels. At the application level, defining optimization settings—such as editable regions, deformation ranges, and design-preservation constraints—typically relies on manual translation of domain expertise by experienced engineers, making the process difficult to automate and weakly coupled with subsequent solvers. At the technical level, surrogate-based optimization methods often suffer from limited reliability when applied to heterogeneous geometry databases or when optimization trajectories enter out-of-distribution (OOD) regions. These issues stem from three primary sources of error in surrogate-assisted inverse design: (1) local bias arising from representing heterogeneous design families with a single global surrogate; (2) insufficient coverage of the query geometry by the historical simulation database; and (3) extrapolation errors induced by the optimization process itself moving candidates into sparsely sampled regions.

Methodology

The authors propose a knowledge-constrained shape-optimization framework integrated with a Mixture-of-Experts Neural Operator (MoE-NO) to enable high-confidence design. The framework operates through two tightly coupled modules:

1. Knowledge-Constrained Design-Space Construction
This module translates domain knowledge ( $K$ ) into quantifiable parameters for deformation operators.

Knowledge Retrieval and Grounding: Using Retrieval-Augmented Generation (RAG) over design standards and guidelines, and a Vision-Language Model (VLM) for visual analysis, the system identifies editable regions, admissible deformation directions, and preservation constraints (e.g., packaging, safety, manufacturability).
3D Control Box Reconstruction: The VLM localizes semantic editable regions in multi-view images, which are then fused into a set of three-dimensional editable control boxes ( $C$ ).
Constraint-Aware Deformation: A Direct Free-Form Deformation (DFFD) operator converts these control boxes and retrieved bounds into smooth local geometry updates ( $G'(\theta)$ ). This ensures the numerical optimizer searches within a knowledge-constrained space $\Omega(C)$ rather than an unconstrained perturbation space, preserving design identity and functional constraints.

2. High-Confidence Optimization with MoE-NO
This module addresses surrogate reliability through a physics-solver-in-the-loop strategy.

MoE-NO Surrogate: Instead of a single global model, the surrogate employs a gating network to route geometries to multiple expert neural operators. This allows the model to capture distinct geometry–performance relations across heterogeneous datasets (e.g., MPV, SUV, Sedan).
OOD Detection: Before optimization, the encoder of the MoE-NO maps query geometries to a latent space. Using nearest-neighbor Euclidean distance, KNN mean distance, and cosine similarity, the system determines if a query is OOD relative to the historical database.
- If OOD: The system triggers a "distribution-level gate," generating local high-fidelity samples around the query, adding a new expert branch to the MoE-NO, and fine-tuning the model to adapt to the new design family.
Uncertainty-Gated Online Refinement: During optimization, a Mahalanobis-distance-based uncertainty score is computed in the latent space for candidate optima.
- If uncertainty is low: The candidate is accepted without a high-fidelity solver call.
- If uncertainty is high: The candidate is sent to a high-fidelity CFD solver. If the solver result deviates significantly from the surrogate prediction or fails to confirm improvement, the system triggers "online refinement." This involves generating local samples near the candidate, updating the training database, and fine-tuning the MoE-NO before restarting the optimization.

Key Contributions

Automated Knowledge Integration: A framework that automatically translates unstructured engineering knowledge and user intent into structured, editable control boxes and deformation bounds using LLMs, VLMs, and RAG, reducing reliance on manual engineering setup.
MoE-NO Architecture: A Mixture-of-Experts Neural Operator that improves surrogate accuracy and trend prediction on heterogeneous aerodynamic datasets by dynamically routing inputs to specialized experts.
Dual-Gate Reliability Strategy: A novel optimization loop that combines distribution-level OOD detection (to handle new design families) and optimization-level uncertainty gating (to handle search-induced extrapolation). This strategy selectively uses expensive high-fidelity solvers only for database enrichment and local error correction, rather than for every candidate evaluation.

Results

The framework was evaluated on in-house datasets for MPV, SUV, and Sedan vehicles, with a held-out "Sedantest" geometry serving as an OOD case.

Surrogate Accuracy: On the in-distribution test set, MoE-NO achieved a Mean Absolute Percentage Error (MAPE) of 1.16% and a trend-prediction accuracy of 94.34%, outperforming baseline models Transolver (MAPE 1.52%, trend 90.34%) and DragSolver.
In-Distribution Optimization: CFD-validated drag coefficient reductions of 10.52% (MPV) and 4.17% (Sedan) were achieved without online refinement. For the SUV case, where the initial surrogate prediction had high uncertainty, online refinement reduced the surrogate-CFD discrepancy from 2.78% to 1.25% and improved the validated drag reduction from 4.36% to 7.96%.
Out-of-Distribution Adaptation: For the held-out Sedantest geometry (initially OOD with $R^2 = -18.42$ and MAPE = 14.88%), the distribution-level gate triggered the addition of a new expert and fine-tuning. This restored predictive capability ( $R^2 = 0.9525$ , MAPE = 1.84%) and enabled a CFD-validated drag reduction of 7.43%.

Significance

The paper claims that this framework provides a practical route to high-confidence design in scenarios where both physical evaluation and manual design-space construction are costly. By converting engineering rules into explicit optimization constraints and using uncertainty quantification to selectively invoke high-fidelity solvers, the method balances computational efficiency with reliability. It addresses the limitations of purely data-driven approaches by ensuring that optimization trajectories remain within physically and industrially valid regions while adapting to new design families when necessary. The authors note that the approach is currently limited to single-objective drag reduction and relies on the quality of retrieved knowledge and visual grounding, suggesting future work in multidisciplinary optimization and variable-fidelity feedback.